.png)
Earlier this year, we introduced LTX-2, a multimodal AI model capable of generating synchronized audio and video with high fidelity and long sequence consistency, and committed to releasing it as open source. Today, we’re following through on that commitment: the complete LTX-2 model, including the full model weights, a distilled version, several LoRAs, and more, is now publicly available.
High-fidelity multimodal models rarely ship as open weights, and when they do, they’re often difficult to run, hard to extend, or not suitable for real creative or production workflows. LTX-2 breaks that pattern. It delivers production-grade audio–video generation, strong creative control, and real-world efficiency — running on modern consumer GPUs and integrating directly into the tools creators rely on today.
We believe strongly that the future of creative AI should be open, collaborative, and hands-on. This release gives developers, researchers, and studios everything they need to run, experiment with, fine-tune, and build on top of the same creative engine that powers our own products.
What Exactly Are We Releasing?
This release makes a full LTX-2 development stack publicly available:
Model Weights: Both the main model and a distilled version, enabling local inference, fine-tuning, and production use across a range of hardware.
LoRA Adapters: A set of LoRAs that demonstrate how LTX-2 can be adapted to specific styles or tasks, and serve as templates for building your own.
Trainer: A modular training framework that includes multimodal data pipelines, reference configurations, multi-GPU support, and evaluation tools, making it straightforward to fine-tune or extend the model for real creative workflows.
LTX-2 has also been optimized using NVIDIA’s AI and GPU acceleration tooling to run efficiently across the entire RTX ecosystem - from consumer RTX 5090 GPUs to enterprise-grade DGX-9 systems.
What LTX-2 Can Do
LTX-2 is the engine behind our creative ecosystem, built to deliver the fidelity, control, and efficiency real workflows demand. The model supports:
High-Fidelity Generation: Up to 20 seconds of synchronized audio and video, rendered at native 4K resolution and up to 50 fps. This is a first for open-weight audio-video generation.
Accurate AV Sync: Expressive lip sync, coherent dialogue, and detailed ambient audio generated in a single pass.
Motion Realism: Dynamic scenes with stable motion, consistent identity, and frame-to-frame coherence.
Fine-Grained Control: Multi-keyframe conditioning, LoRA-based customization, 3D camera logic, and multimodal inputs (text, image, video, audio, depth) for precise creative direction.
Radical Efficiency: A hyper-efficient latent space and step-distilled architecture that achieve comparable results with significantly less compute. LTX-2 runs efficiently on high-end consumer GPUs, making high-fidelity multimodal generation accessible without specialized hardware.
This combination of fidelity and efficiency is essential for real creative work. Fast feedback enables iteration; iteration enables discovery. LTX-2 is designed for that loop.
Why We’re Committed to Open Source
Across the industry, the trend is clear: models are getting larger, access is narrowing, and more of the creative stack is being locked behind paywalled APIs. We believe in the opposite approachL openness, transparency, and local control. This will create the breakthroughs that matter most.
High-quality open models are rare, and open models capable of production-grade results are rarer still. With LTX-2, we’re choosing to make a frontier-level system fully inspectable and adaptable, because we believe the next leap in creative intelligence will come from broad participation, not closed walls.
Here’s why this matters:
Supporting Research: Open weights let researchers explore new architectures, training strategies, and multimodal approaches. LTX-2 can serve as a solid starting point for entirely new lines of work.
Unlocking New Use Cases and Extensions: Developers can fine-tune LTX-2 for their own domains, build LoRAs, create distillations, and design creative workflows that address real-world needs.
Strengthening a Global Community: When people start creating adapters, tools, demos, integrations, and research using LTX-2, the model gets more useful for everyone. Open ecosystems get stronger and spark more innovation over time.
Our goal is to contribute a strong foundation that the community can meaningfully build upon. In the long arc of our vision, LTX-2 is the open model stage of a system that will learn from real creative workflows and evolve into an AI-driven rendering engine, and eventually, a form of creative intelligence.
How to Get Started
Download the Model: Access the weights, code, and benchmarks on GitHub and Hugging Face. https://github.com/Lightricks/LTX-2 and https://huggingface.co/Lightricks/LTX-2
Try the Tools: Use LTX-2 through ComfyUI or your own custom inference setup. https://github.com/Lightricks/ComfyUI-LTXVideo
Join the Community: Connect with other developers, share feedback, and collaborate on extensions. https://discord.gg/ltxplatform
The Future Is Creative. And Open.
By releasing LTX-2 openly, we’re making it a shared foundation for multimodal creativity. We expect the most exciting progress to come from how people fine-tune it, adapt it, distill it, and build it into new creative workflows. Whether through custom LoRAs, new workflows, or entirely novel applications, the community will take LTX-2 in directions we haven’t imagined - and that evolution is exactly why we wanted to open it in the first place.
We’re excited to see where you take it next!
